Key takeaway: AI-powered BPMN modeling can quickly turn a textual process description into a clear proposed model. However, the generated diagram does not prove that the process is correct from a business perspective, handles every exception, or is directly executable. This requires process owner approval, BPMN expertise, and technical validation.
In this installment of “BPMN 2.0 in Practice,” we use a purchase request example to demonstrate how to create a model from text. We also cover validation and explain how business users can move from a textual description to an executable workflow in xFLOWer.ai.
Practical focus: This series is not a complete BPMN 2.0 reference manual—the official notation set is far broader, and here we focus on the most essential elements and practical workflows used in everyday operations.
1. A good model starts with a precise process description
AI works with the information provided. If the description says “the manager approves the request, then the purchase is made,” it does not explain what happens if the request is rejected, information is missing, or the manager is unavailable. In such cases, the system may omit important branches or fill gaps in the process with assumptions.
An assumption is not a business rule. Therefore, separate confirmed information, open questions, and operational details to be decided later in the input. The description should include at least the following:
- Trigger and objective: what triggers the process, and what outcomes can it produce?
- Roles: who submits, checks, approves, and fulfills the request?
- Decision rules: which monetary threshold or condition changes the route?
- Exceptions: how do we handle rejection, requests for missing information, and missed deadlines?
- Data: which fields, documents, and data from external systems are required?
For example: an employee submits a request, and the purchasing officer checks it for completeness and returns it for correction if information is missing. In this example, financial approval is also required above 500 000 forints. This is an illustrative rule, not a general procurement requirement.
2. Which BPMN elements should the text become?
Model generation is not about arranging sentences into boxes. The meaning of the description must be expressed through BPMN elements. It is particularly important to distinguish between performing work, waiting, making decisions, and assigning responsibility.
| BPMN element | What does it represent? | Procurement example |
|---|---|---|
| Event | Something that starts a process, occurs during it, or brings it to an end. | Request received; waiting period expires; request closed. |
| Activity | Work to be performed by a person or a machine. | Checking a request; recording a purchase order. |
| Exclusive gateway | Selects a single outgoing branch based on a condition. | Is the request approved or rejected? |
| Parallel gateway | When splitting, it activates every outgoing branch. When merging, it waits for an arrival from every incoming branch. | Legal and financial reviews start independently of each other. |
| Event-based gateway | The next event to occur determines how the process continues. | A supplier response arrives, or the response deadline expires. |
| Lane | A division of responsibilities or organizational structure within a participant’s process. | Tasks of the requester, procurement officer and finance team. |
Choose the gateway based on the process logic. If a known data value determines the path, an exclusive gateway may be needed. If every branch must be executed, use a parallel gateway. If we do not yet know which event will occur first, use an event-based gateway.
A lane is not the same as an access permission: placing a task in the finance lane does not remove the need to configure access separately in the execution system. Communication between separate participants should be represented by message flows, not sequence flows crossing pool boundaries.
3. From text to model: six verifiable steps
- Define the process boundaries. Start with a specific process variant. Decide, for example, whether the model will cover the process up to request approval or through to completion of fulfillment.
- Create a structured description. Specify the participants, steps, conditions, deadlines and exceptions. Mark uncertain details as open questions.
- Ask for clarifying questions first. Have the AI list missing rules and inconsistencies. This allows you to clarify, before drawing the diagram, who makes the decision when an amount changes, for example.
- Generate a proposed model. Request descriptive task names, explicit decision conditions and a list of assumptions. If necessary, also specify the required export format.
- Validate using specific cases. Walk through standard requests, rejected requests, requests requiring additional information and overdue requests. The next step must be clear in every case.
- Record the approved version. Document the model version, agreed rules, reviewer and any outstanding technical questions.
For example, a useful instruction could be: “Create a proposed BPMN 2.0 model from the description. Do not invent approval rules. First list the questions requiring clarification, then the elements and their relationships. Identify your assumptions separately.”
The result may be a structured list of elements or—depending on the tool’s capabilities—BPMN XML. Successfully importing the XML does not, by itself, confirm that the process works correctly.
4. Expert review: ten common sources of error
The AI-generated model should be reviewed from three perspectives: business correctness, BPMN semantics and executability. The process owner reviews the rules, the modeler checks the notation, and the implementer examines the technical requirements.
- Unclear start: exactly which event creates a new process instance?
- Missing completion: do rejection and withdrawal, as well as approval, lead to a well-defined end state?
- Incorrect gateway type: are alternative and parallel execution kept distinct?
- Overlapping conditions: do the rules for the exclusive split unambiguously determine which path is taken?
- Unhandled decision case: is there a path for every possible value, or a justified default branch?
- Incorrect synchronization: does the join avoid waiting for a branch that was never activated?
- Mixed-up connections: are sequence flows and message flows used in the appropriate places?
- Overlooked exception: is the handling of requests for additional information, technical errors and timeouts clearly defined?
- Unassigned task: can a responsible person and a substitution rule be assigned to every human activity?
- Endless rework: is there a controlled way to move forward from or terminate the correction loop?
Write short test cases for validation: input data, expected path and expected outcome. A “high-value, incomplete request,” for example, tests both the handling of missing information and threshold-based approval. This is more valuable than simply inspecting the diagram visually.
5. The Role of xFLOWer.ai and the Requirements for Execution
In xFLOWer.ai, business users can independently create an executable procurement workflow from a textual description, complete with a request form and approval rules. The example in this article is worth testing with two requests. At exactly 500 000 forints, the process skips financial approval; at 500 001 forints, it assigns a task to Finance. Before going live, the process owner checks whether this rule complies with company policy and whether the task actually reaches the correct approver.
Importing and exporting BPMN XML and executing it on another engine are separate capabilities that must be clarified according to the specific requirements.
What Else Is Needed for a Working Process?
Alongside the BPMN diagram, you also need a data model, forms, task assignment, permissions, integrations and error handling. For example, you must define where the supplier identifier comes from, what happens when a system call fails, and which data changes trigger a new approval.
The set of BPMN elements supported by the execution environment must also be checked. Not every standard construct can run in every system. Integration tests, business acceptance testing and logging checks are required before going live. Only authorized data should be submitted for AI processing; remove sensitive information from examples wherever possible.
6. Related Articles and the Next Step
The other articles in this series explore individual aspects of validated modeling:
- BPMN 2.0 Elements Explained with Examples: using events, activities, gateways and lanes correctly.
- Exclusive, Parallel and Event-Based Gateways: When Should You Use Each? Distinguishing between decisions, parallel execution and waiting for events.
- The 10 Most Common BPMN Modeling Mistakes: identifiable issues and guidance for correcting them.
- BPMN vs. Flowcharts vs. UML Activity Diagrams: choosing a notation that suits the modeling objective.
- How Does a BPMN Diagram Become an Executable Process? Connecting data, integrations and execution rules.
Summary: the value of AI-based BPMN modeling lies in faster, verifiable preparation. Good results require precise input, deliberate element selection, documented expert review and a tested implementation.
Would You Like to Try It on Your Own Process? Bring a short process description and a few exception scenarios to a consultation with an xFLOWer expert. In xFLOWer.ai, these can become an executable workflow that you can test with specific requests before going live.
This article is part of our complete guide: BPMN 2.0 in Practice: A Complete Guide from Modeling to a Working Process
xFLOWer.ai
Design the process behind this article – for free!
AI reads the processes hidden in this article; open any of them in xFLOWer.ai with one click.
